VULTRITION: Vulnerability Dataset Nutrition Labels

vultrition is a lightweight Python command-line tool for generating nutrition-label style summaries for function-level vulnerability datasets. It turns a dataset and a TOML configuration file into machine-readable JSON facts and an SVG label that can be included in dataset READMEs, benchmark reports, papers, or dataset catalogs.

VULTRITION does not rank datasets or certify their quality. Instead, it exposes facts that are often hidden, incomplete, or reported inconsistently: dataset size, metadata completeness, vulnerability-type coverage, class imbalance, duplicate or near-duplicate functions, train/test contamination, and structural characteristics of the code.

Core contributions

• Compact dataset nutrition labels: Produce standardized numerical summaries for function-level vulnerability datasets instead of long, hard-to-compare documentation.
• Reproducible fact extraction: Use a TOML config to describe dataset metadata, file locations, split information, and field mappings.
• Quality and structure in one view: Report facts about completeness, balance, CWE/project coverage, CVE timespan, uniqueness, split contamination, lines of code, token counts, and cyclomatic complexity.
• Benchmark transparency: Help researchers choose datasets, document dataset versions, and interpret vulnerability-detection results beyond headline model scores.

Installation

Install using pip:

pyhon3 -m pip install vultrition

Install from source:

python3 -m pip install -e .

Python 3.11 or newer is recommended. On older Python versions, install tomli for TOML support.

If runtime dependencies are not installed automatically, install them with:

python3 -m pip install tomli svgwrite lizard tiktoken pygments tqdm numpy torch transformers faiss-cpu

For GPU-enabled FAISS, install a FAISS build that matches your CUDA environment instead of faiss-cpu.

Quick start

Create a configuration template:

vultrition --create-config-template vds-config.toml

Edit the generated TOML file so that the dataset metadata, file paths, and field names match your dataset.

Run the analysis pipeline:

vultrition --config vds-config.toml --run_analysis --output vds_nutrition_label_data.json

Generate the SVG nutrition label:

vultrition --create_vultrition_label vds_nutrition_label_data.json --svg_output vultrition_label.svg

Input data

VULTRITION supports vulnerability datasets stored as:

• .csv
• .json
• .jsonl
• .ndjson

A dataset can be provided either as predefined train, test, and optional valid splits, or as a single data file. JSON files may contain a top-level list or an object with a data, records, or samples list.

Configuration

VULTRITION expects a TOML file with a top-level [dataset] section. The fields section maps the column names in your dataset to the canonical fields used by the analysis.

[dataset]
name = "Example Vulnerability Dataset"
description = "Function-level vulnerability dataset"
version = "1.0.0"
license = "MIT"
has_runable_code_or_test_cases = false
languages = "c,c++"

[dataset.files]
train = "data/train.jsonl"
test = "data/test.jsonl"
valid = "data/valid.jsonl"
# For a single-file dataset, use this instead of the split paths:
# data = "data/all.jsonl"

[dataset.fields]
function = "func"
label = "label"
vuln_label_value = 1
cve = "cve"
cwe = "cwe"
project = "project_url"

[dataset.analysis.quality_metrics]
completeness = true
diversity = true
balance = true
timespan = true
uniqueness = true
cross_contamination = true

[dataset.analysis.structural_metrics]
loc = true
tokens = true
cyclomatic_complexity = true

Field mapping

Config key	Meaning
function	Source code snippet or function body.
label	Original vulnerability label.
vuln_label_value	Label value that should be interpreted as vulnerable. Matching records are normalized to 1; all others to 0.
cve	CVE identifier field. Used for completeness and timespan analysis.
cwe	CWE identifier field. Strings, delimited strings, and lists are supported.
project	Project, repository, or source identifier.

Reported facts

Quality facts

VULTRITION reports facts that help characterize dataset composition and reliability:

• number of functions
• metadata completeness
• vulnerable/non-vulnerable balance
• unique projects and CWEs
• CVE-year timespan
• nearest-neighbor similarity and near-duplicate rate
• cross-contamination between train, validation, and test splits when splits are available

Similarity-based facts are computed with code embeddings and FAISS nearest-neighbor search. The default near-duplicate threshold is 0.95.

Structural facts

VULTRITION also reports structural properties of the analyzed functions:

• lines of code
• token counts
• cyclomatic complexity

These facts are useful when comparing datasets for modern code models, where long or complex functions may affect training and evaluation.

CLI reference

Option	Description
-c, --config	Path to a VULTRITION TOML config file.
--create-config-template, --create_config_template	Create a template config file at the provided path or inside the provided directory.
-o, --output	JSON output path for analysis results. Defaults to vds_nutrition_label_data.json.
--run_analysis	Run the full analysis pipeline.
--create_vultrition_label	Create an SVG label from an existing JSON result file.
--svg_output	SVG output path. Defaults to vultrition_label.svg.
-v, --verbose	Print parsed config details and sample records during analysis.
--version	Print the installed CLI version.

Outputs

Running analysis produces a JSON file with dataset metadata, quality facts, and structural facts:

vultrition --config vds-config.toml --run_analysis --output results.json

The JSON file can be rendered as an SVG label:

vultrition --create_vultrition_label results.json --svg_output vultrition_label.svg

Video Demo

Video demo is available on YouTube

Notes

• The generated nutrional labels for all 15 datasets can be found in examples/labels/.
• The raw results can be found in results/.

Contributing

Contributions, bug reports, and feature requests are welcome. Please open an issue or submit a pull request on the GitHub repository.